Gemini 4 Neon from Google is now available on AI Gateway. Gemini 4 Neon is a major upgrade over the earlier, lighter Gemini 4 model, delivering output quality close to Google’s larger Gemini 4 models with faster responses. It’s built for co…
Firehose
Filtered to tagged “natural language processing” · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives
Browse by tag
GPT-6.2 Astra, a model from OpenAI, has achieved a perfect score of 100% on the ARC-AGI-3 benchmark, surpassing its predecessor GPT-6 Astra. This achievement marks the first time a model has scored 100% on ARC-AGI-3, with GPT-6.2 Astra using 64.1% fewer actions than the human baseline. AI summary
TrnRT is an open-source LLM inference engine for AWS Trainium, utilizing NKI megakernels, 1M-token context, and an OpenAI-compatible API. It serves a single trn2.48xlarge model, runs decode as NKI megakernels, and maintains a prefix cache across device and host memory. TrnRT is built on top of the AgentX results, serving GLM-5.3-Flash on one Trainium2 box with NKI layer megakernels. AI summary
Qwen4-Max from Alibaba is now available on AI Gateway. It is the first model in the Qwen4 series: a reasoning model with native vision and a 1M-token context window, built for coding, professional work, and long-running agent tasks. Qwen4-M…
XYBench is a benchmark of 8K+ questions that tests how large language models respond to requests based on misconceptions, measuring their ability to identify and address the underlying goal. The benchmark evaluates models' ability to recognize and redirect users toward effective solutions, focusing on pragmatic emphasis and misconception identification. The results show that current models often prioritize the stated request over the underlying goal, highlighting a need for improvement in model evaluation and training methods. AI summary
Researchers at NeuraCryption have developed SPECTRA, a neural model compression framework that can compress LLMs by 2.80x to 3.31x in under 50 MB RAM, achieving near-lossless compression and real-time generation speeds. The framework is based on Shannon source coding theory and continuous rate-distortion bounds, and eliminates empirical calibration passes by analytically determining parameter sensitivity from learned normalization scales. AI summary